Observability has long been a core challenge in distributed systems. By collecting logs, metrics, and traces, engineers gain visibility into system behavior. In traditional microservices, execution is request-driven, topology is mostly static, and sampling is commonly used to manage data volume. Observability data is typically analyzed using SQL-based analytics engines.

AI agents fundamentally change what traces represent. Instead of predefined service paths, traces capture dynamic runtime execution: model invocations, tool calls, and orchestration steps selected based on context. It is impossible to predict execution paths in advance, and each request may produce a unique, data-dependent execution graph.

This shift introduces two key challenges: rapidly increasing data scale and continuously evolving execution graphs that cannot be modeled ahead of time. We show how Apache Iceberg, combined with a graph engine, provides a flexible and cost-efficient foundation for scalable observability in AI agent–driven systems.